Genome-Wide Association Study

A genome-wide association study (GWAS) is a hypothesis-free, population-scale statistical genetics study design that scans millions of common and low-frequency single-nucleotide polymorphisms (SNPs) and small insertions/deletions across the entire DNA genome in thousands to millions of individuals to identify genetic variants associated with complex quantitative traits or binary disease phenotypes. Introduced conceptually by Neil Risch and Kathleen Merikangas (1996) and executed at scale beginning with the landmark 2005 age-related macular degeneration (CFH) study by Robert Klein et al. and the 2007 Wellcome Trust Case Control Consortium (WTCCC) study of 14,000 cases across seven common diseases, GWAS replaced low-powered candidate-gene studies and linkage mapping in pedigrees with genome-wide allelic association across unrelated or distantly related biobank cohorts.

Because human complex traits—such as standing height, blood pressure, low-density lipoprotein (LDL) cholesterol, type 2 diabetes, schizophrenia, and coronary artery disease—are polygenic (conforming to Ronald Fisher's 1918 infinitesimal model, in which hundreds or thousands of genetic loci each contribute small additive effects alongside environmental variance), modern GWAS biobanks (such as the UK Biobank, All of Us, FinnGen, and Biobank Japan) combine SNP microarray genotyping and haplotype imputation with linear and logistic mixed-effects models (BOLT-LMM, SAIGE, REGENIE). Downstream of locus discovery, GWAS summary statistics serve as the empirical input for Linkage Disequilibrium Score Regression (LDSC) to estimate SNP heritability (h²_SNP), statistical fine-mapping and colocalization to pinpoint causal non-coding variants within topologically associating domains, two-sample Mendelian randomization to test causal epidemiological risk factors, and Polygenic Risk Scores (PRS) for clinical risk stratification.

Genotype Calling, Imputation, and Additive Dosage (0, 1, 2)

In a typical GWAS cohort of N individuals (i = 1, ..., N) assayed across M biallelic genetic variants (j = 1, ..., M, where M ≈ 10⁷ after imputation against whole-genome reference panels such as TOPMed, Haplotype Reference Consortium, or 1000 Genomes using hidden Markov chain Li-Stephens haplotype copying), each SNP j has a reference allele A₁ and an alternative (effect) allele A₂ with population minor allele frequency (MAF) p_j = min(f(A₂), 1 − f(A₂)).

Under the standard additive genetic model, the raw genotype G_ij ∈ {0, 1, 2} counts the number of copies of the effect allele A₂ carried by individual i at variant j (0 for homozygous A₁A₁, 1 for heterozygous A₁A₂, and 2 for homozygous A₂A₂). For imputed variants, G_ij ∈ [0, 2] represents the posterior expected allele dosage G_ij = P(A₁A₂) + 2P(A₂A₂), filtered by an imputation quality information metric INFO ≥ 0.30 to 0.80. Under Hardy-Weinberg Equilibrium (HWE, tested in controls via Pearson's χ² test, typically requiring P_HWE > 10⁻⁶), the genotypic variance at locus j is Var(G_j) = 2 p_j(1 − p_j), and the standardized genotype matrix X has entries:

X_ij = ( G_ij − 2 p_j ) / [ 2 p_j(1 − p_j) ]^1/2

Marginal Single-Variant Regression and the Genome-Wide Significance Threshold

For a continuous quantitative trait y = (y₁, ..., y_N)^T (such as rank-based inverse-normal transformed BMI or height), classical fixed-effect GWAS fits a separate marginal linear regression model at each variant j, adjusting for a covariate matrix C (age, sex, genotyping batch, and the top 10 to 20 ancestry Principal Components, PCs computed by EIGENSTRAT to control for population stratification):

y_i = μ + β_j G_ij + γ^T c_i + ε_i, where ε_i ~ N(0, σ_e²)

For a binary case-control disease phenotype y_i ∈ {0, 1}, GWAS fits a logistic regression model on the log-odds scale:

ln[ P(y_i = 1) / (1 − P(y_i = 1)) ] = μ + β_j G_ij + γ^T c_i

where exp(β̂_j) is the per-allele Odds Ratio (OR). Testing the null hypothesis H₀: β_j = 0 via the Wald statistic z_j = β̂_j / SE(β̂_j) or score test yields the two-sided asymptotic p-value P_j = 2 Φ(−|z_j|), where z_j² ~ χ₁² under H₀.

Because millions of correlated SNPs are tested simultaneously across the genome, evaluating significance at α = 0.05 without multiple-testing correction would produce hundreds of thousands of false positives. Empirical permutation analyses in European and Cosmopolitan populations (Pe'er et al., 2008; Dudbridge & Gusnanto, 2008) established that the common human genome (MAF ≥ 1%) behaves statistically as approximately M_eff ≈ 1,000,000 (10⁶) independent linkage disequilibrium blocks. Applying a family-wise error rate (FWER) Bonferroni correction of α = 0.05 across 10⁶ independent tests establishes the universal genome-wide significance threshold:

P < 0.05 / 10⁶ = 5 × 10⁻⁸ (corresponding to |z| ≥ 5.451 or χ₁² ≥ 29.72, plotted as −log₁₀(P) ≥ 7.301 on a Manhattan plot)

For whole-genome sequencing studies that additionally test rare coding and non-coding variants down to MAF ≥ 0.01%, the effective number of independent tests increases to ~5 × 10⁶–10⁷, shifting the stringent threshold to P < 5 × 10⁻⁹.

In biobanks of N = 500,000+ participants, thousands of pairs of individuals share cryptic 2nd- to 5th-degree relatedness, and Conditioning on a random polygenic background substantially reduces residual phenotypic variance—boosting statistical power to detect new loci. Linear Mixed Models (LMMs) model the phenotype vector y as the sum of fixed effects (the candidate SNP j on chromosome c plus covariates C), a random polygenic genetic effect g drawn from a covariance matrix proportional to the Genetic Relationship Matrix (GRM, K), and residual environmental noise e:

y = X_j β_j + C γ + g + e, where g ~ N(0, σ_g² K_LOCO), e ~ N(0, σ_e² I_N), and K = (1 / M_GRM) X X^T

To avoid proximal contamination—the loss of power that occurs if the candidate SNP j (or SNPs in LD with it) is included in both the fixed effect and the random GRM K—modern mixed-model engines construct K_LOCO using a Leave-One-Chromosome-Out (LOCO) scheme that excludes all variants on the chromosome carrying SNP j.

Because explicit eigendecomposition of an N × N GRM costs O(N³) time (~1.25 × 10¹⁷ operations for N = 500,000), three scalable architectures were developed:

1. BOLT-LMM (Loh et al., 2015): Avoids forming K in memory by using preconditioned conjugate gradient (PCG) iteration directly on genotype matrix-vector products X(X^Tv) and fits a non-infinitesimal Bayesian Gaussian mixture (spike-and-slab) prior on SNP effects, increasing effective sample size by up to 15–25% for highly heritable quantitative traits.
2. SAIGE (Zhou et al., 2018): When a binary disease has severe case-control imbalance (e.g., 500 cases vs. 400,000 controls) and a variant has low minor allele frequency (MAF < 1%), the normal approximation of the score test statistic fails in the extreme tails, generating massive false-positive rates. SAIGE fits a Generalized Linear Mixed Model (GLMM) via average-information restricted maximum likelihood (AI-REML) and replaces the normal tail approximation with a cumulant-generating function Saddlepoint Approximation (SPA), accurately calibrating p-values down to P = 10⁻¹² even at 1:1,000 case-control ratios.
3. REGENIE (Mbatchou et al., 2021): Decomposes mixed modeling into a two-step machine-learning ridge regression pipeline: Step 1 partitions ~500,000 genotyped SNPs into contiguous blocks of 1,000 variants, computes ridge-regression predictors across a grid of shrinkage parameters, and combines them via a second-level cross-validated ridge regression (or logistic ridge regression) to produce 22 LOCO polygenic predictions ŷ_LOCO. Step 2 subtracts ŷ_LOCO as an offset and tests all M = 50,000,000 imputed variants in parallel using Firth's penalized likelihood or SPA.

Distinguishing Polygenicity from Stratification via LDSC

Before the biobank era, systematic inflation of genome-wide test statistics was monitored using the genomic control inflation factor (λ_GC) (Devlin & Roeder, 1999), defined as the ratio of the empirical median of the observed χ₁² test statistics across all genome-wide SNPs to the theoretical median of a central χ₁² distribution with 1 degree of freedom (0.4549364):

λ_GC = median(z₁², z₂², ..., z_M²) / 0.4549

In an underpowered study of N = 2,000 samples, λ_GC > 1.05 warns of uncorrected population stratification or cryptic relatedness. However, in a well-powered biobank GWAS (N = 500,000) of a highly polygenic trait such as height or educational attainment, true causal variants are distributed across the majority of LD blocks in the genome, pushing λ_GC to 1.30–2.50 even in the complete absence of population stratification.

In 2015, Brendan Bulik-Sullivan, Hilary Finucane, Benjamin Neale, Alkes Price, and colleagues resolved this confound by formulating Linkage Disequilibrium Score Regression (LDSC). Define the LD Score (ℓ_j) of SNP j as the sum of its squared correlation coefficients with all neighboring variants k within a 1 centimorgan (1 Mb) window: ℓ_j = ∑_k r_jk². Because a SNP with a high LD score tags more neighboring variants, it has a proportionally higher probability of tagging a true polygenic causal effect, whereas pure population stratification (a) inflates all SNPs equally regardless of their LD score:

E[ χ_j² | ℓ_j ] = ( N · h²_SNP / M ) · ℓ_j + N · a + 1

Regressing the observed GWAS χ_j² statistics against the precomputed LD scores ℓ_j cleanly separates the two effects:

• The LDSC Regression Slope (N h²_SNP / M): Estimates the total common SNP-heritability (h²_SNP)—the proportion of phenotypic variance explained additively by all common genotyped/imputed SNPs (helping quantify how much of twin heritability H² is captured by common variants vs. rare variants under purifying selection).
• The LDSC Intercept (N a + 1): Isolates residual confounding from population stratification and cryptic relatedness. For instance, a height GWAS with λ_GC = 1.95 and an LDSC intercept of 1.03 indicates that (1.95 − 1.03) / (1.95 − 1.00) ≈ 96.8% of the test-statistic inflation represents true polygenic biological signal rather than confounding. Extending LDSC to cross-trait products z_1jz_2j estimates the genome-wide genetic correlation (r_g) between any two traits from summary statistics alone without requiring overlapping individuals.

Polygenic Risk Scores (PRS) and Cross-Ancestry Portability

By aggregating the small additive effects of thousands to millions of variants across the genome, a Polygenic Risk Score (PRS, or Polygenic Score PGS) predicts an individual i's genetic liability toward a complex disease or trait:

PRS_i = ∑_j=1..M ŵ_j G_ij

Because marginal GWAS effect estimates β̂_j include redundant correlated signal across LD blocks, constructing an optimal weight vector ŵ requires either heuristic Linkage Disequilibrium Clumping and P-value Thresholding (C+T / PRSice) or Bayesian inference shrinkage of the marginal effects using the LD matrix (LDpred2, PRS-CS continuous shrinkage priors, or SBayesR). In cardiovascular medicine and oncology, individuals in the top 1% to 5% tail of a genome-wide PRS for coronary artery disease or breast cancer carry a 3-fold to 5-fold elevated lifetime disease risk—comparable in magnitude to monogenic familial hypercholesterolemia (LDLR) or moderate-penetrance germline mutations. However, because LD patterns, allele frequencies, and tagging correlations diverge across continental populations due to genetic drift after the Out-of-Africa bottleneck, a PRS trained strictly on European-ancestry biobanks suffers an approximately 50% to 75% attenuation in predictive accuracy (R²) when applied to African-ancestry cohorts—motivating multi-ancestry GWAS discovery and trans-ancestry Bayesian PRS methods (PRS-CSx) combined with statistical fine-mapping and colocalization to weight true shared causal variants.

1. Wellcome Trust Case Control Consortium. "Genome-wide association study of 14,000 cases of seven common diseases and 3,000 shared controls." Nature 447, 661–678 (2007). https://doi.org/10.1038/nature05911
2. Bulik-Sullivan, B. K., et al. "LD Score regression distinguishes confounding from polygenicity in genome-wide association studies." Nature Genetics 47, 291–295 (2015). https://doi.org/10.1038/ng.3211
3. Loh, P. R., et al. "Efficient Bayesian mixed-model analysis increases association power in large cohorts." Nature Genetics 47, 284–290 (2015). https://doi.org/10.1038/ng.3190
4. Zhou, W., et al. "Efficiently controlling for case-control imbalance and sample relatedness in large-scale genetic association studies." Nature Genetics 50, 1335–1341 (2018). https://doi.org/10.1038/s41588-018-0184-y